[Mesa-dev] [PATCH 08/10] radeonsi: wait for outstanding memory instructions in TCS barriers

[Marek Olšák]

From: Marek Olšák <marek.olsak at amd.com>

Cc: 13.0 12.0 <mesa-stable at lists.freedesktop.org>
---
 src/gallium/drivers/radeonsi/si_shader.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index b914efb..45896bd 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -3176,28 +3176,32 @@ static void build_type_name_for_intr(
 }
 
 static void build_tex_intrinsic(const struct lp_build_tgsi_action *action,
 				struct lp_build_tgsi_context *bld_base,
 				struct lp_build_emit_data *emit_data);
 
 /* Prevent optimizations (at least of memory accesses) across the current
  * point in the program by emitting empty inline assembly that is marked as
  * having side effects.
  */
+#if 0 /* unused currently */
 static void emit_optimization_barrier(struct si_shader_context *ctx)
 {
 	LLVMBuilderRef builder = ctx->gallivm.builder;
 	LLVMTypeRef ftype = LLVMFunctionType(ctx->voidt, NULL, 0, false);
 	LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, "", "", true, false);
 	LLVMBuildCall(builder, inlineasm, NULL, 0, "");
 }
+#endif
 
+/* Combine these with & instead of |. */
+#define LGKM_CNT 0x07f
 #define VM_CNT 0xf70
 
 static void emit_waitcnt(struct si_shader_context *ctx, unsigned simm16)
 {
 	struct gallivm_state *gallivm = &ctx->gallivm;
 	LLVMBuilderRef builder = gallivm->builder;
 	LLVMValueRef args[1] = {
 		lp_build_const_int32(gallivm, simm16)
 	};
 	lp_build_intrinsic(builder, "llvm.amdgcn.s.waitcnt",
@@ -5228,21 +5232,21 @@ static void si_llvm_emit_barrier(const struct lp_build_tgsi_action *action,
 				 struct lp_build_tgsi_context *bld_base,
 				 struct lp_build_emit_data *emit_data)
 {
 	struct si_shader_context *ctx = si_shader_context(bld_base);
 	struct gallivm_state *gallivm = bld_base->base.gallivm;
 
 	/* The real barrier instruction isn’t needed, because an entire patch
 	 * always fits into a single wave.
 	 */
 	if (ctx->type == PIPE_SHADER_TESS_CTRL) {
-		emit_optimization_barrier(ctx);
+		emit_waitcnt(ctx, LGKM_CNT & VM_CNT);
 		return;
 	}
 
 	lp_build_intrinsic(gallivm->builder,
 			   HAVE_LLVM >= 0x0309 ? "llvm.amdgcn.s.barrier"
 					       : "llvm.AMDGPU.barrier.local",
 			   ctx->voidt, NULL, 0, 0);
 }
 
 static const struct lp_build_tgsi_action tex_action = {
-- 
2.7.4